Fast Anther Dehiscence Status Recognition System Establishing by Deep Learning to Screen Heat Tolerant Cotton

Zhihao Tan, Jiawei Shi, Rongjie Lv, Qingyuan Li, Jing Yang, Yizan Ma, Yanlong Li, Yuanlong Wu, Rui Zhang, Huanhuan Ma, Yawei Li, Zhu Li, Longfu Zhu, Xianlong Zhang, Jie Kong, Wanneng Yang, Ling MinResearch 2021 · DOI

TL;DR

Researchers have developed a deep learning-based system to automate the identification of cotton anther dehiscence—the process where pollen-producing organs open up. By ensembling improved Faster R-CNN models, the system achieves high-accuracy recognition that can replace tedious manual counting, providing a high-throughput tool for breeding heat-tolerant cotton in a warming climate.

What Problem Does This Paper Address?

Cotton is a global economic powerhouse, but its yield is highly sensitive to heat stress. The "canary in the coal mine" for cotton fertility is the anther. Under normal conditions, anthers dehisce (split open) to release pollen for fertilization. However, high temperatures (HT) often cause anthers to remain closed (indehiscent), leading to male sterility and crop failure.

Traditionally, breeders have had to manually count dehiscent vs. indehiscent anthers to identify heat-tolerant varieties. This is a nightmare of a task: anthers are tiny, hundreds are packed into a single flower, and they often overlap or blend into the background. This manual process is slow, subjective, and nearly impossible to scale for the large populations required in modern genetic research.

Methods

The authors compared and optimized two popular deep learning architectures to see which could best handle the "small object detection" challenge of cotton anthers:

  1. YOLOv5 (You Only Look Once): A single-stage detector chosen for its speed and potential for mobile deployment. To handle overlapping anthers, the team implemented Soft-NMS (Non-Maximum Suppression), which decays the confidence scores of overlapping boxes rather than just deleting them.
  2. Faster R-CNN: A two-stage detector known for higher accuracy on small objects. The authors pushed this model further by adding:
  3. Feature Pyramid Networks (FPN): To better capture multi-scale features.
  4. Advanced Data Augmentation: Techniques like Mixup and Cutmix to help the model generalize and handle sample imbalances.
  5. Model Ensembling: The final "production" system isn't just one model; it’s an ensemble of four different Faster R-CNN variations, with their results averaged to minimize error.

Datasets

The study utilized a robust, multi-year dataset:

  • Raw Data: 38,895 high-definition RGB images of cotton anthers.
  • Collection: Data was gathered between 2016 and 2019 across three distinct geographic locations in China (Wuhan, Turpan, and Alar) to ensure environmental diversity.
  • Annotated Set: 2,845 images were manually labeled with "open" and "close" tags using LabelImg software.
  • Benchmark: The dataset was split 7:3 for training and validation.

Key Insights and Ideas

  • Accuracy vs. Mobility Trade-off: While YOLOv5 is the "speed king" and can be deployed on a smartphone for real-time field use, it struggled with quantitative precision. Faster R-CNN, though heavier, proved that for scientific phenotyping where every anther counts, the two-stage approach is superior.
  • The Power of the Ensemble: The authors found that while individual improvements (like FPN or Multi-Scale training) helped, the real "aha moment" came from ensembling. The ensemble model outperformed any single model, reaching an $R^2$ (coefficient of determination) high enough to officially replace manual counting.
  • Biological Validation: The AI didn't just find boxes; it found biology. The system successfully quantified that heat stress drops the average anther dehiscence rate from ~84% down to ~35%, and it identified specific lines (S003 and S004) that maintained high fertility under stress.

Key Results

  • Performance: The ensembled Faster R-CNN model achieved an $R^2$ of 0.8765 for "open" anthers and 0.8539 for "close" anthers.
  • Detection Metric: The YOLOv5 model achieved an [email protected]:0.95 of 0.485, slightly higher than the base Faster R-CNN, but it fell significantly behind in the $R^2$ correlation required for accurate counting.
  • Throughput: The system can process batch inputs of images rapidly, allowing researchers to screen 30+ varieties and identify heat-tolerant germplasms in a fraction of the time it takes a human.

Limitations and Future Work

The authors acknowledge that while dehiscence is a primary marker, it isn't the only one. Future iterations of the system could look at:

  • Spatial Phenotypes: Measuring the growth position of anthers and the physical distance between anthers and stigmas, which also affects pollination.
  • Cross-Crop Application: Adapting the model to detect anther states in other economically important crops like rice or maize.
  • Platform Expansion: Moving the high-accuracy ensemble models onto cloud servers to allow field researchers to upload photos and receive instant, high-precision reports.

Summary updated

Abstract

Abstract Background Cotton is one of the most economically important crops in the world. The fertility of male reproductive organs is a key determinant of cotton yield. The anther dehiscence or indehiscence directly determine the probability of fertilization in cotton. Thus, the rapid and accurate identification of cotton anther dehiscence status is important for judging anther growth status and promoting genetic breeding research. The development of computer vision technology and the advent of big data have prompted the application of deep learning techniques to agricultural phenotype research. Therefore, two deep learning models (Faster R-CNN and YOLOv5) were proposed to detect the number and dehiscence status of anthers. Result The single-stage model based on YOLOv5 has higher recognition efficiency and the ability to deploy to the mobile end. Breeding researchers can apply this model to terminals to achieve a more intuitive understanding of cotton anther dehiscence status. Moreover, three improvement strategies of Faster R-CNN model were proposed, the improved model has higher detection accuracy than YOLOv5 model. We have made four improvements to the Faster R-CNN model and after the ensemble of the four models, R2 of “open” reaches 0.8765, R2 of “close” reaches 0.8539, R2 of “all” reaches 0.8481, higher than the prediction result of either model alone, and can completely replace the manual counting method. We can use this model to quickly extract the dehiscence rate of cotton anther under high temperature (HT) condition. In addition, the percentage of dehiscent anther of randomly selected 30 cotton varieties were observed from cotton population under normal conditions and HT conditions through the ensemble of Faster R-CNN model and manual observation. The result showed HT varying decreased the percentage of dehiscent anther in different cotton lines, consistent with the manual method. Conclusions The deep learning technology first time been applied to cotton anther dehiscence status recognition instead of manual method to quickly screen the HT tolerant cotton varieties and can help to explore the key genetic improvement genes in the future, promote cotton breeding and improvement.
BibTeX
@inproceedings{tan2021fast,
  author = {Zhihao Tan and Jiawei Shi and Rongjie Lv and Qingyuan Li and Jing Yang and Yizan Ma and Yanlong Li and Yuanlong Wu and Rui Zhang and Huanhuan Ma and Yawei Li and Zhu Li and Longfu Zhu and Xianlong Zhang and Jie Kong and Wanneng Yang and Ling Min},
  title = {{Fast Anther Dehiscence Status Recognition System Establishing by Deep Learning to Screen Heat Tolerant Cotton}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-1128527/v1},
}
Figure